01
Upload
POST /api/upload
Client guards the 5 MB cap; server re-verifies size, MIME, and the %PDF magic bytes. Upstash holds a per-IP fixed window — 3 uploads per day, no exceptions.
≤ 5 MB≤ 50 pages3 / IP / day
Architecture · Live demo
Behind the curtain
Six steps, one cited answer.
A live look at the pipeline you just used — the embeddings, the vector store, the cron that quietly throws everything out after a day. What you see here is what's actually running in production.
Active documents
—
Embedded chunks
—
Avg / document
—
Next cleanup
—
Pipeline
How a question becomes a cited answer
02
Chunk
pdf-parse · per page
pdf-parse walks the document one page at a time. Each page is normalised and sliced into ~500-token blocks with 50-token overlap, snapping to sentence boundaries when it can.
≈ 500 tok / chunk50-tok overlapsoft breaks
03
Embed
Vertex · text-embedding-004
Chunks pack into batches under Vertex's 20k-token limit, pessimistically estimated at chars/2. Retries on 429 with exponential backoff, plus a 250 ms breath between batches.
768-dim vectors8k tok / batch5× retry
04
Store
Supabase · pgvector
Vectors land in a dedicated rag schema with RLS enabled (anon role bounced at the door). An ivfflat cosine index handles approximate-nearest-neighbour search.
schema ragivfflat · lists=100RLS on
05
Ask
POST /api/chat
The question is embedded once, fed to the match_chunks RPC scoped to the active document. Top 5 passages return as context, each keeping its page number for citations.
top 5 · cosinedoc-scoped20 / IP / day
06
Stream
Vertex · Gemini 2.5 Flash
A system prompt enforces 'cite [Source N] or refuse'. Tokens stream back as Server-Sent Events. The client renders citation footnotes inline as the text arrives.
max 1024 tokT=0.2SSE
Auto-cleanup
The 24-hour eviction
Schedule
—
Retention
—
cascade delete
Next run in
—
A Vercel cron hits /api/cleanup daily with a Bearer token. The endpoint calls rag.cleanup_old_documents('24 hours') — a single SQL function that drops anything older than the cutoff and lets the foreign-key cascade do the rest. No filename logs are kept.
Live data
Currently in the demo
Filenames are redacted — fellow visitors uploaded these, and that's their business.
| Label | ID | Pages | Chunks | Age | Deletes in |
|---|---|---|---|---|---|
| Loading… | |||||
Stack
Specification
FrameworkNext.js 14 · App Router · TS
StylingTailwind CSS
GenerationVertex AI · Gemini 2.5 Flash
EmbeddingsVertex AI · text-embedding-004 · 768d
Vector storeSupabase Postgres + pgvector
Rate limitUpstash Redis · fixed window
PDF viewerreact-pdf · client-side highlights
Local cacheIndexedDB blobs · idb wrapper
HostVercel · daily cron
Sourcegithub.com/chitai-dev (private demo)
For hire
Want this on your stack?
One-time build. I scope it with you, build it on your GCP project, hand over the code and the keys. No retainer, no monthly fee, no “tier” you have to stay subscribed to. The infrastructure is yours from day one — I just got it shipped.
What's included
- Production-grade build of this exact pattern, hardened for your data
- Deployed to your GCP project — you own billing, models, infrastructure
- Hybrid retrieval (BM25 + dense + reranker) and a proper eval harness
- VPC / CMEK / VPC-SC if you need it. Auth (IAP, OIDC) wired through
- Observability — Cloud Logging, Trace, per-tenant cost dashboards
- Handover doc + a runbook your ops team can actually use
What I don't do
- Monthly retainers or seat-based pricing
- Lock-in via a hosted layer you can't remove
- Black-box pipelines you'd have to rebuild to extend
- Fluffy 'AI strategy' decks
Scope a project
Send me the rough idea — document count, target users, deadline — and I'll come back with a fixed scope, fixed price, and a build window.